Self-Sensing Technology of Rotor Displacement for Six-Pole Radial Active Magnetic Bearing Based on Physics-Informed Neural Network

To address the high cost and complexity of displacement sensors in magnetic bearing systems, a prediction model for rotor displacement based on a physics-informed neural network (PINN) with adaptive adjustment of multiple loss function weights is proposed for inverter-driven six-pole radial active magnetic bearings. First, the mathematical model of the levitation force is derived using the Maxwell tensor method, and the system's left invertibility and observability are proven, validating the feasibility of using a neural network for displacement prediction. Then, dynamic weight adjustment loss function and boundary constraint layers are incorporated into the PINN model to enable self-sensing of rotor displacement. Through comparison tests with other displacement prediction models, the results show that the PINN displacement prediction model proposed in this paper performs the best in terms of prediction accuracy. Subsequently, a simulation system is built, and levitation and anti-interference tests are performed. The simulation results demonstrate that the PINN model exhibits high fitting accuracy and robust anti-interference performance. Finally, validation experiments are conducted on an experimental platform, confirming the effectiveness and feasibility of the proposed method.